Skip to content
Insights/Article

Same-Named Customers and Similar Products: How Should AI Clarify the Target Before Acting?

Key takeaway

Confirm the target before acting; do not auto-select the closest or most recent. List candidates with distinguishing fields such as city, contact, tax ID, or model, and ask the user to choose or add information. Carry the entity ID into the task. High-impact actions always confirm; routine lookups may interrupt less.

AI listing two same-named customer candidates with city and contact fields before executing a quote task

Consider a hypothetical case. Sales tells AI: "Send a quote to Huaxin at the last price." Two customers contain "Huaxin" in their names: one in Shanghai, one in Suzhou. AI picks the Shanghai company contacted most recently, generates a quote, and sends it. Two days later, procurement in Suzhou calls to ask where the quote is, and sales discovers the mistake.

AI did not invent a customer. In an ambiguous instruction, it made a choice it had no authority to make. Identical names, similar models, and overlapping abbreviations are common in business data. Confirming the target before acting is the most easily skipped step in an AI workflow, with a very direct consequence.

"Closest" is not "the one"

Auto-selecting by recent contact, name similarity, or order value looks smart but hides business judgment inside the model. The recently contacted company may not be the one meant; the highest similarity score may be a duplicate record created last week.

The safer principle: ask when ambiguous, and carry the entity ID even when unambiguous. Ask with minimal interruption, but never substitute a guess for confirmation.

Candidates and distinguishing fields

A name alone is useless in a candidate list. Show fields that distinguish:

  • Customer: city, contact, last four digits of tax ID, latest order, customer ID.
  • Product: full model name, version, specification difference, applicable region.
  • Contact: company, title, last four phone digits, email domain.
  • Order: order number, date, amount, status.

Choose fields the user can judge without opening a detail page. Two to four are usually enough. Avoid sensitive data; a tax ID can show the last four digits only.

The confirmation screen

Guideline G10 recommends disambiguation or reduced scope when the user goal is uncertain.

The screen lets the user choose, not watch AI explain itself. Common elements:

  • Two to five candidates, each with distinguishing fields.
  • An exit: "None of these; I will add information."
  • After selection, show the entity ID and key fields before execution.
  • For high-impact actions, a second "Send/Execute" button after confirmation.

If there are more than five candidates, the user's description is too vague. Ask for more information instead of dumping a long list.

When confirmation is mandatory

Mandatory: sending quotes, sending email, changing or cancelling orders, refunds, address changes, creating contracts, adjusting permissions. These affect external parties or have irreversible results.

Recommended: batch changes, cross-customer operations, actions involving money or stock.

May not interrupt: read-only lookups. For "show Huaxin's orders," return both companies' summaries and let the user inspect, rather than asking first.

Not interrupting does not mean not labeling. Lookup results show entity IDs and distinguishing fields so the user can spot a wrong match.

After confirmation: carry the entity ID

When the user chooses "Huaxin (Shanghai)," the task uses the customer ID, not name matching again. Names repeat; IDs do not.

For multi-step tasks—generate quote, send email, create follow-up—every step references the same entity ID. Re-matching by name midway can pick wrong again.

The execution record stores the entity ID and how it was confirmed: explicit user choice, unique system match, or resolved after added information. When something goes wrong, the confirmation point is recoverable.

A hypothetical clarification dialogue

Turn

Content

User

Send a quote to Huaxin at the last price.

AI

Two customers contain "Huaxin": Huaxin (Shanghai), contact Mr. Zhang; Huaxin (Suzhou), contact Mr. Li. Which one?

User

Shanghai.

AI

Confirmed: Huaxin (Shanghai), customer ID C-1024. A quote will be generated from the latest one and sent to Mr. Zhang. Send?

User

Confirm.

The dialogue costs two extra turns and prevents one misdirected send. Low-risk lookups can skip confirmation; send-type actions justify the two turns.

Entity confirmation in batch tasks

Batch imports, bulk price changes, and bulk notifications cannot skip confirmation. Every row carries an entity ID, and before submission the system shows the match result: how many rows matched uniquely, how many have several candidates, and how many matched nothing.

Do not auto-select the first candidate for a multi-match row; mark it for a person. Do not create a new customer or product for an unmatched row; first check whether the name is misspelled or the record was never created.

The batch preview shows the actual names and IDs of the first few rows so a person can see whether they are the intended set. After execution, report how many succeeded, skipped, and failed, and let failed rows be downloaded and corrected separately.

The worst batch outcome is not an error message; it is "everything looked successful" while several rows changed the wrong entity. The preview and the result receipt exist to prevent exactly that. Keep the batch ID and the confirmed entity list with the execution record, so a later correction can identify affected rows and the person who approved the batch. Large batches should run in chunks with a checkpoint after each; a failed chunk does not roll back successful ones unless that is explicitly designed.

Confirm the Target Before Taking Action

[AI task entity confirmation checklist]

1. Ambiguity detection
- Can the system detect identical, similar, and duplicate records:
- Does it avoid auto-selecting by recency or similarity:
- Does it ask for more information when candidates exceed five:

2. Candidate display
- Do candidates show city, contact, ID, and other distinguishing fields:
- Is full sensitive data avoided:
- Is there a "none of these" exit:

3. Execution
- Do high-impact actions require confirmation:
- Does execution use the entity ID rather than the name:
- Do multi-step tasks reference the same ID throughout:

4. Records
- Are the confirmation method and confirmer recorded:
- Can results trace back to the confirmation point:
- Is there a recall or correction flow for mis-sends:

Pick the entity type most often misselected—same-named customers or similar models—draw the confirmation screen, and have three salespeople try it. If they must open another system to tell candidates apart, the fields are insufficient. Ticket classification and handoff are neighboring problems, while entity confirmation happens before execution. Action permissions are covered in AI agent permissions, approval, and human takeover; responsibility after an error is in who is accountable when an AI agent errs. The scope of AI task execution and confirmation can be discussed within AI applications and workflows.

After choosing the target, keep queries within read-only boundaries. For similar product names, also review model and version identities.

Sources

Let's talk about your project

Websites / Business systems / AI applications

Chat on WhatsApp